Predicting multimorbidity longitudinal course with probabilistic graphical models

Marcos L. P. Bueno1,2, Carla Silva2, Mariana Lobo2, Pedro Pereira Rodrigues2

1Radboud University Nijmegen (NL), 2Faculdade de Medicina, CINTESIS, Universidade do Porto (PT)

OBJECTIVE: understanding multimorbidity interaction in patients over time is challenging. In particular, we aim at identifying multimorbidity factors associated with patients admitted with an acute myocardial infarction.

METHODS: We considered a cohort with 28 co-morbidities from 500 patients admitted with acute myocardial infarction in Portuguese hospitals (2011-2015), totaling 851 hospitalizations. For modeling, we considered several probabilistic graphical models that naturally allow to represent multiple variables together. These models, which include dynamic Bayesian networks and hidden Markov models, are used to uncover correlations that take time into account with different temporal abstractions. Bootstrapping is used to estimate the significance of associations found in the graphical models.

RESULTS: Preliminary results using dynamic Bayesian networks suggest that direct correlations between diseases appear to be sparse, which means that predicting a disease needs only a few other predictors. Further results include testing the significance of such correlations, analyzing the correlation between the number of hospitalizations and diseases, among others.

keywords: administrative data, multimorbidity, Bayesian networks, acute myocardial infarction

Presentation: Predicting multimorbidity longitudinal course with probabilistic graphical models